Change in the API ?

[Optimox]

I am using the latest timm version '1.0.7'

If I run this piece of code everything runs as expected:

import timm
network = timm.create_model(model_name="efficientnet_b0", #efficientnet_b0, resnet18
                                pretrained=True,
                                num_classes=0,
                                in_chans=3,
                                drop_rate=0,  # default : 0
                                drop_path_rate=0,
                                global_pool="",  # default: 'avg'
                                )


a = torch.ones((2, 3, 128, 128))
out = network(a)

assert(out.shape[1]==network.num_features)

However if I switch efficientnet_b0 to resnet18 I get an error as out.shape = [2, 64, 32, 32]

This used to work previously, did something change in the create_model API recently? What is now the recommended way to extract features of any model and get the corresponding size automatically ?

[rwightman]

@Optimox the API did change a little bit, I added a 'head_hidden_size' for determining pre_logits shape for models where that's different, however it should not impact efficientnet or resnet (head_hidden_size=num_features there) and I just checked and it seems fine, model.num_features is 512 for resnet18 and 1280 for efficiientnet_b0 and it matches the output

#2224 has cut and paste code of what the latest tests are validating for the models

EDIT: also, if what your observing happens, the unit tests would fail as that's checked as per the snippests in that other discussion, no other changes are made to the lib?

[Optimox]

I ran pip install --upgrade --force-reinstall timm=="1.0.7" and it does run fine now... That is strange, it looks like there is a combo which creates a silent bug.

Thank you very much for your help @rwightman!

[rwightman]

@Optimox hmm, that is odd, but taking note in case someone else ends up with that sort of failure on upgrading.